Awesome
<p align="center"> <h2 align="center"><strong>FlowCon: Out-of-Distribution Detection using Flow-Based Contrastive Learning</strong></h2> <h3 align="center"><span style="font-size:1em;" color><strong>ECCV 2024</strong></span> </h3> <p align="center"> <a href="https://saandeepa93.github.io/"><strong> Saandeep Aathreya</strong></a>, <a href="https://scanavan.github.io/"><strong> Shaun Canavan</strong></a> <br> <span style="font-size:1em; "><strong> University of South Florida, USA</strong>.</span> <br> </p> </p> <p align="center"> <a href="https://arxiv.org/abs/2407.03489" target='_blank'> <img src="https://img.shields.io/badge/arXiv-Paper-greem.svg"> </a> </p>:bulb: Contributions:
<p align="center"> <img src="docs/static//images/contrib.png" width="410" /> </p> <!--  -->-
A new density-based OOD detection technique called FlowCon is proposed. We introduce a new loss function $L_{con}$ which contrastively learns class separability in the probability distribution space. This learning occurs without any external OOD dataset and it operates on fixed classifiers.
-
The proposed method is evaluated on various metrics - FPR95, AUROC, AUPR-Success, and AUPR-Error and compared against state of the art. We observe that FlowCon is competitive or outperforms most methods under different OOD conditions. Additionally, FlowCon is stable even for a large number of classes and shows improvement for high-dimensional features
-
Histogram plots are detailed along with unified manifold approximations (UMAP) embeddings of the trained FlowCon model to respectively showcase it’s OOD detection and class-preserving capabilities. We also show FlowCon’s discriminative capabilities.
Results
<!-- ### **FAR-OOD likelihood plots when $D_{in}=CIFAR10$ on ResNet-18 and WideResNet models** <p align="center"> <img src="docs/static/images/ll_cifar10.png" width="410" height="400" /> </p> ### **FAR-OOD likelihood plots when $D_{in}=CIFAR100$ on ResNet-18 and WideResNet models** <p align="center"> <img src="docs/static/images/ll_cifar100.png" width="410" height="400" /> </p> ## **FlowCon Comparison for Semantic and Covariate Shift** --> </br> <p align="center"> <img src="docs/static/images/main_res.png" width="400" height="350"/> </p> <div style="display: flex; justify-content: space-between;"> <img src="docs/static/images/far_near.png" alt="Image 1" style="width: 45%;"/> <img src="docs/static/images/near.png" alt="Image 2" style="width: 45%; "/> </div> <!-- <p align="center"> <img src="docs/static/images/far_near.png" width="310" height="200"/> </p> <!-- <p align="center"> <img src="docs/static/images/near.png" width="310" height="200"/> </p> --> <!-- ## **UMAP Embeddings for Semantic and Covariate Shift** <p align="center"> <img src="docs/static/images/umap.png" width="410" height="200"/> </p> <p align="center"> <img src="docs/static/images/umap_cifar100.png" width="410" height="200"/> </p> -->Configurations
- All config files are present under path
./configs/experiments - We have 4 config files
cifar10_3.yaml -> CIFAR10 with ResNet18
cifar10_5.yaml -> CIFAR10 with WideResNet
cifar100_6.yaml -> CIFAR100 with ResNet18
cifar100_4.yaml -> CIFAR100 with WideResNet
Pretrained Classifiers
- The original ResNet18 and WideResNet classifiers are taken from Heatmap paper
Training FlowCon model
CIFAR-10
Training Resnet and WideResNet FlowCon model for CIFAR10. Follow similar commands for other configs
CUDA_VISIBLE_DEVICES=0 python ./trainer/train_flow.py --config cifar10_3 #Resnet
CUDA_VISIBLE_DEVICES=0 python ./trainer/train_flow.py --config cifar10_5 #WRN
CIFAR-100
CUDA_VISIBLE_DEVICES=0 python ./trainer/train_flow.py --config cifar100_6
CUDA_VISIBLE_DEVICES=0 python ./trainer/train_flow.py --config cifar100_4
Getting OOD Scores
- Inference follows similar arguments
CUDA_VISIBLE_DEVICES=0 python ./tester/emperical_params.py --config cifar10_5
CUDA_VISIBLE_DEVICES=0 python ./tester/emperical_params.py --config cifar100_4
-
The $k$ distributions will be saved under
./data/distributionsdirectory -
The results will be saved under
./data/resultsdirectory
Pre-trained model weights
All the pretrained classifiers, and pretrained FlowCon models can be found in link